GPT Live 1 Livekit Inference Not Working?

Getting Started
1

Hello,

I see in Livekit 1.8.8 Python Livekit Inference is supposed to Support GPT Live but I keep getting HTTP 503.

2

I just tested with the attached, and it seems to work:

agent.py (5.8 KB)

There havenā€™t been any outages that could explain the 503, which I assume would have been an issue between our backend and OpenAI.

Did you specify inference.DuplexModel ?

3

I tested your full agent.py unchanged using livekit-agents==1.8.8 and our LiveKit project credentials. The agent connected to the room and received test audio, but inference.DuplexModel(model="openai/gpt-live-1") failed to connect after retries.

I also captured the underlying handshake response:

{ā€œcodeā€:ā€œmodel_unavailableā€,ā€œmessageā€:ā€œrequested gpt-live model is unavailableā€,ā€œtypeā€:ā€œserver_errorā€}

With the same credentials, openai/gpt-realtime-2.1 connects successfully. Weā€™re using the default production inference endpoint with no override.

Failing request ID: dbd33686-eee5-4b55-a19a-8929e189c943
Timestamp: October 19, 2026, at 15:57:19 UTC.

4

Could it be a ZDR HIPAA thing that is enabled on my account?

5

Good catch, itā€™s HIPAA related.

Your workspace is marked as HIPAA required, but I checked the code as the source of truth and the model is not served to HIPAA-compliant workspaces. The ā€˜officialā€™ list is HIPAA Eligible Products and Services | LiveKit . I see internal discussions that the team are working on this, but please donā€™t take that as a guarantee the model will at some stage be allowed on HIPAA-complaint workspaces, though obviously we would like it to.

Iā€™m more concerned about your confusing error message - I thought we had a dedicated error for this, something like ā€˜model will not be served to a HIPAA-complaint projectā€™. Let me look into that separately.

Note that all LiveKit Inference models are ZDR, including this one, so this was not related to ZDR.

6

Yes I think a better error code would be helpful. Was banging my head trying to figure out what I was doing wrong! Hopefully will be compliant soon!

7

Non-compliant pipeline models DO give a helpful error, itā€™s just realtime models that do not (hence my confusion earlier). So, we need to address that.